Concrete crack detection is essential for ensuring infrastructure safety and integrity. Although deep learning techniques have achieved success in this domain, their high computational demands limit real-time deployment on resource-constrained devices. To address this challenge, we compared lightweight convolutional neural network (CNN) architectures for the concrete crack classification problem, including EfficientNet B0, NASNet-Mobile, MobileNetV3Small, ResNet50V2, and DenseNet121. In our comparative experiments, we evaluated the CNN-based models using 5-fold cross-validation on the Concrete Crack Images for Classification (CCIC) dataset. The models achieved accuracies of up to 99\% with approximately 5–20 million parameters, making them well-suited for deployment on resource-constrained devices. Furthermore, our analysis confirms that these models remain robust under diverse environmental conditions. Overall, these lightweight CNN architectures provide a scalable and efficient solution for real-time concrete crack detection, paving the way for proactive infrastructure maintenance and enhanced structural health monitoring.